Causal Thinking Lance J. Rips1 Northwestern University To appear in: J. E. Adler & L. J. Rips (Eds.), Reasoning: Studies of Human Inference and Its Foundation. Cambridge, UK: Cambridge University Press. Causal Thinking / 2 Causal Thinking One damn thing leads to another. I forget to open the garage door this morning, back my car into the door, and splinter it. The actions we perform cause other events—my backing up causes the splintering. But events of other kinds—nonactions—have their effects too. With no help from me, last night’s storm caused a branch to fall from a tree, putting a hole in my roof. Much as we might like to forget them, we often keep track of events like these and the causes that unite them. Although we might not have predicted these events, we can remember and reconstruct part of the causal sequences after they occur. In retelling the events of last summer, for example, we tend to relate the events in forward causal order, starting, say, at the beginning of our trip to Virginia in May and proceeding chronologically. If we want to mention other kinds of events from the same period, such as our summer work experiences, we may start again at the beginning of the summer, moving along the events in a parallel causal stream (Barsalou, 1988). We also remember fictional stories in terms of the causal changes that compose their main plot line, remembering less about events falling on deadend side plots (Trabasso & Sperry, 1985). We sometimes attribute causal powers to concrete objects as well as to events, but we can understand this sort of talk as an abbreviation for event causation. If Fred caused the glass to break that’s because one of Fred’s actions—maybe his dropping it—caused the breaking. I’ll take event causation as basic in this article on the strength of such paraphrases. We remember causes and effects for event types as well as for event tokens. Ramming heavy objects into more fragile ones typically causes the fragile items damage; repeating phone numbers four or five times typically causes us to remember them for awhile. Negotiating routine events (e.g., Schank & Abelson, 1977), constructing explanations (e.g., Lewis, 1986), and making predictions all require memory for causal relations among event categories. Causal generalities underlie our concepts of natural kinds, like daisies and diamonds (e.g., Ahn & Kim, 2000; Barton & Komatsu, 1989; Gelman & Wellman, 1991; Keil, 1989; Rehder & Hastie, 2001; Rips, 1989, 2001) and support our concepts of artifacts like pianos or Causal Thinking / 3 prisms. Our knowledge of how beliefs and desires cause actions in other people props up our own social activities (e.g., Wellman, 1990). The importance of causality is no news. Neither are the psychological facts that we attribute causes to events, remember the causes later, and reason about them—although, as usual, controversy surrounds the details of these mental activities. Recently, though, psychologists seem to be converging on a framework for causal knowledge, prompted by earlier work in computer science and philosophy. Rhetorical pressure seems to be rising to new levels among cognitive psychologists working in this area: For example, “until recently no one has been able to frame the problem [of causality]; the discussion of causality was largely based on a framework developed in the 18th century. But that’s changed. Great new ideas about how to represent causal systems and how to learn and reason about them have been developed by philosophers, statisticians, and computer scientists” (Sloman, 2005, p. vii). And at a psychological level, “we argue that these kinds of representations [of children’s knowledge of causal structure] and learning mechanisms can be perspicuously understood in terms of the normative mathematical formalism of directed graphical causal models, more commonly known as Bayes nets… This formalism provides a natural way of representing causal structure, and it provides powerful tools for accurate prediction and effective intervention” (Gopnik et al., 2004, p. 4). It’s a little unfair to catch these authors in mid rhetorical flight. But the claims for these formalisms do provoke questions about how far they take us beyond the simple conclusions I’ve already mentioned. Kids and adults learn, remember, and apply causal facts. As a card-carrying CP (i.e., cognitive psychology) member, I believe that kids and adults therefore mentally represent these facts. But what’s new here that further illuminates cognitive theorizing? Here’s the gloomy picture: The new methods are at heart data-analytic procedures for summarizing or approximating a bunch of correlations. In this respect, they’re a bit like factor analysis and a whole lot like structural equation modeling. (If you think it surprising that psychologists should seize on a statistical procedure as a model for ordinary causal thinking, consider that another prominent theory in this area is Kelley’s, 1967, ANOVA model; see the section on Causation from Correlation, below, and Gigerenzer, 1991.) The idea that people use these Causal Thinking / 4 methods to induce and represent causality flies in the face of evidence suggesting that people aren’t much good at normatively correct statistical computations of this sort (e.g., Tversky & Kahneman, 1980). Off hand, it’s much more likely that what people have are fragmentary and error-prone representations of what causes what. The rosier picture is the one about “great new ideas.” The jury is still out, and I won’t be resolving this issue here. But sorting out the claims for the new causal representations highlights some important questions about the nature of causal thinking. How are Causal Relations Given to Us? Here’s a sketch of how a CD player works (according to Macaulay, 1988): A motor rotates a spindle that rotates the CD. As the CD turns, a laser sends a beam of light through a set of mirrors and lenses onto the CD’s surface. The light beam lands on a track composed of reflecting and nonreflecting segments that have been burned onto the CD. The reflecting segments bounce the light beam back to a photodiode that registers a digital “on” signal; the nonreflecting segments don’t bounce the light back and represent an “off” signal. The pattern of digital signals is then converted into a stereo electrical signal for playback. You could remember this information in something like the form I just gave you—an unexciting little narrative about CD players. But the new psychological approach to causal knowledge favors directed graphs like Figure 1 as mental representations—“causal maps” of the environment (Gopnik et al., 2004). This graph contains nodes that stand for event types (e.g., the CD player’s motor rotating or not rotating, the CD turning or not turning) and directed links that stand for causal connections between these events (the motor rotating causes the CD’s turning; the laser producing a beam and the mirror-lens assembly focusing the beam jointly cause the beam to hit the CD’s surface). Of course, no one disputes the fact that people can remember some of the information these diagrams embody. Although people can be overconfident about their knowledge of mechanical devices like this one (Rozenblit & Keil, 2002), Causal Thinking / 5 they’re nevertheless capable of learning, say, that the CD player’s motor causes the CD to turn. What’s not so clear is how they acquire this cause-effect information, how they put the component facts together, and how they make inferences from such facts. In this section we’ll consider the acquisition problem, deferring issues of representation and inference till the second part of this article. ---------------------------------------------------------------------------------------- Please insert Figure 1 about here. ----------------------------------------------------------------------------------------- Causation in Perception You’re not likely to get much of the information in Figure 1 by passively observing a CD player, unless you already know about the nature of similar devices. But sometimes you do get an impression of cause from seeing objects move. Repeated sightings of an event of type E1 followed by an event of type E2 may provide evidence that E1 causes E2. Rather weak evidence, but evidence nonetheless. When we later see an example of the same sequence, we can infer the causal link. But psychologists sometimes claim there is a more intimate perception of cause in which an observer directly experiences one event causing another. Perceptual studies. In a famous series of demonstrations, Michotte (1963) rigged a display in which a square appeared to move toward a second square and to stop abruptly when they touched. If the second square then began to move within a fixed interval of the touching and at a speed similar to that of the first square, observers reported the first square causing the second to move, launching it. Michotte’s extensive experiments aimed to isolate the purely perceptual conditions that produce this immediate impression of causality, but there’s a paradoxical quality to his efforts. The first square in the display doesn’t actually cause the second to move. The displays showed 2-D projections of simple geometrical forms whose movements could be carefully controlled behind the scenes. (In those days before lab computers, Michotte engagingly created his displays using striped disks rotating behind slits or using pairs of moving slide projectors.) The goal was therefore not to determine when people correctly detect causal relations in their environment but instead to uncover the cues that lead them to report Causal Thinking / 6 causality.2 Michotte himself discusses a number of situations in which people report one event causing another, even though the interaction is physically unlikely or impossible. In one such case, a square A moves at 30 cm/s and comes into contact with another square B, which is already moving at 15 cm/s.